1. Field of the Invention
The present invention relates to an image processing apparatus and method for detecting a specific object pattern from an image.
2. Description of the Related Art
Image processing methods for automatically detecting specific object patterns from images are very useful and are used in, for example, determination of the faces of persons. Such methods are available in many fields including communication meetings, man-machine interfaces, security, monitor systems for tracing the faces of persons, and image compression. Various technologies for detecting faces from images are described in M.h. Yang, D. J. Kriegman, and N. Ahuja “Detecting Faces In Images: A Survey” IEEE Trans. On PAMI, Vol. 24, No. 1, pp. 34-58, January, 2002. In particular, an AdaBoost-based method described in P. Viola and M. Jones “Robust Real-time Object Detection” in Proc. of IEEE Workshop SCTV, July, 2001 is widely used in research on face detection because of its high execution speed and detection ratio.
FIG. 8 illustrates an example of a face detector in related art. As illustrated in FIG. 8, the face detector proposed by Viola et al. has a cascade structure in which multiple face identifiers are arranged. The face identifier in each stage determines whether an input image represents a face or non-face, and only the image determined to represent a face proceeds to the next stage. The image reaching the final stage is finally determined to represent a face.
FIG. 9 illustrates exemplary features identified by an identifier in the related art. As illustrated in FIG. 9, in each stage of a cascade structure, many features each belonging to any of four simple features are combined to compose the identifier. Each of the four features corresponds to the difference between the sum of gray-scale values in white rectangles and the sum of gray-scale values in black rectangles. A function for comparing this difference with a threshold value to output “1” or “0” is called weak hypothesis. Several thousands to several tens of thousands of pieces of learning data are used to configure the weak hypothesis. In the learning, one hundred and thirty thousands or more features are generated depending on how the positions and sizes of the rectangles are determined in an image of 24×24 pixels. The AdaBoost algorithm is used to select any of the features.
Although the face detector proposed by Viola, et al. can accurately detect front faces in various illumination conditions because of the enormous amount of learning data, faces subjected to out-of-plane rotations often fail to be detected. In contrast, application of the face detector proposed by Viola, et al. to the upper bodies of persons allows the face detector to function as a person detector that is capable of detecting the objects. However, there are cases where the objects cannot be detected because of the various illumination conditions that are varied.